Indonesian J our nal of Electrical Engineering and Computer Science V ol. 42, No. 2, May 2026, pp. 572 ∼ 583 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v42.i2.pp572-583 ❒ 572 P omelo maturity classication fr om eld-acquir ed images using oil-gland mor phology and a rule-based image-pr ocessing pipeline Sopapun Suwansawang 1 , Harutai Dinsakul 1 , W ir ot Buangam 1 , Jirar oj T osasukul 2 1 Department of Electrical Engineering, F aculty of Science and T echnology , Nakhon P athom Rajabhat Uni v ersity (NPR U), Nakhon P athom, Thailand 2 Department of Mathematics, F aculty of Science, Naresuan Uni v ersity (NU), Phitsanulok, Thailand Article Inf o Article history: Recei v ed Dec 30, 2025 Re vised Feb 24, 2026 Accepted May 26, 2026 K eyw ords: Non-destructi v e classication Oil gland detection Pomelo maturity Rule-based image processing Surf ace feature analysis ABSTRA CT Pomelo maturity assessment in commercial orchards relies predominantly on vi- sual inspection and harv est age records, which introduce inconsistenc y in post- harv est grading. Non-destructi v e al ternati v es such as near -infrared spectroscop y and acoustic sensing ha v e been reported, b ut typically require specialised instru- ments and controlled acquisition conditions. This study in v estig ates the feasibil- ity of oil-gland morphology as an interpretable maturity indicator , implemented as a rule-based image-processing pipeline e x ecutable on standard CPU hardw are without model training. A hierarchical rule-based frame w ork w as de v eloped to classify pomelo maturity from gland count features e xtracted under natural outdoor illumination. Thirty-three Citrus maxima samples (Khao Y ai culti v ar) representing three maturity stages were analysed in this proof-of-concept study ( n = 11 per stage). The pipeline inte grates adapti v e thresholding, subre gion se gmentation, multi-scale morphological dete ction, and threshold-based classi- cation. Detection reliability w as v eri ed on synthetic dot-pattern images prior to real-sample e v aluation. On the collected dataset, the frame w ork achie v ed an o v erall accurac y of 78.8% with a macro-a v eraged F1-score of 0.784. No mis- classication occurred between the immature and mature groups; errors arose e xclusi v ely between adjacent stages. Mean processing time w as 57 seconds per image on a consumer -grade laptop. Gi v en the limited sample size and single- culti v ar scope, these results represent methodological feasibility rather than v al- idated generalisation, and establish a baseline for morphology-based maturity assessment in pomelo. This is an open access article under the CC BY -SA license . Corresponding A uthor: Sopapun Suw ansa w ang Department of Electrical Engineering, F aculty of Science and T echnology Nakhon P athom Rajabhat Uni v ersity 85 Malaiman, Nakhon P athom 43000, Thailand Email: sopapun@webmail.npru.ac.th 1. INTR ODUCTION Field-deplo yable computer vision systems in agriculture must operate under uncontrolled i llumina- tion, heterogeneous backgrounds, and limited computational resources. Deep learning has substantially ad- v anced automated fruit maturity assessment t hrough con v olutional neural netw orks and transformer -based ar - chitectures [1], [2], achie ving strong predicti v e performance across di v erse crops [3]. Ho we v er , these systems J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 573 typically require lar ge labeled datasets, e xtensi v e training, and hardw are capable of real-time inference [4]-[6]. Spectral and h yperspectral imaging methods of fer complem entary biochemical information b ut depend on spe- cialized equipment and calibration procedures [1]. T ogether , these constraints moti v ate continued interest in computationally ef cient, interpretable frame w orks based on con v entional RGB imaging. Pomelo (Citrus maxima) maturity assessment is a representati v e e xample of this challenge. Har v es t timing in commercial orchards is commonly determined through visual inspection and fruit age records, with gro wers relying on subjecti v e surf ace cues including pee l color a n d oil gland appearance. These indicators lack standardized quantitati v e procedures [7], reecting the broader absence of objecti v e, automated systems capable of reliable operation under natural eld conditions. Oil glands embedded in the citrus a v edo layer form discrete surf ace structures that under go de v elop- mental changes during fruit gro wth [8]. Machine vision has been widely applied acr o s s citrus product ion tasks including grading and harv est identication [9], with oil gland surf ace appearance demonstrated as a functional feature for non-destructi v e citrus quality assessment from digital images [10]. F or pomelo specically , the visual contrast between oil glands and surrounding peel tissue has been identied as a discriminati v e feature in multi-parameter maturity e v aluation [7], [11]. Despite this rele v ance, e xplicit spatial quantication of oil- gland morphology within a deterministic, rule-based classication frame w ork has recei v ed limited systematic in v estig ation. As summarized in T able 1, representati v e maturity detection approaches in v olv e trade-of fs among predicti v e capability , computational demand, and deplo yment comple xity . This study in v estig ates the feasibility of using oil-gland morphology as an interpretable maturity indicator through a rule-based image-processing pipeline designed for operation on standard CPU hardw are under practical eld acquisition conditions. The frame w ork inte grates preprocessing, subre gion se gmentation, morphological ltering, and e mpirical decision rules within an OpenCV -based pipeline, emphasizing deterministic decision logic o v er data-intensi v e learning strate gies. T able 1. Comparison of representati v e fruit maturity detection approaches. Characteristics reect intrinsic methodological properties; direct accurac y comparisons across datasets are not intended Approach Computational demand and transparenc y Deplo yment characteristics Spectral imaging [1] Moderate–high; moderate feature interpretability Specialised hardw are; calibration required Deep learning [4], [6] High; limited decision interpretability Strong in controlled settings; lar ge labelled dataset required Proposed rule-based pipeline Lo w-to-moderate; no GPU or specialist hardw are required; e xplicit traceable decision logic Standard RGB camera; platform- independent; no training data required Image acquisition w as performed using consumer -grade RGB de vices, and the processing fra me w ork is platform-independent with no de vice-specic calibration required. The objecti v e is to establish a method- ological baseline for morphology-based pomelo surf ace analysis under practical eld conditions rather than to compete directly with high-capacity data-dri v en models. The main contrib utions are as follo ws: − F ormulation of a morphology-based maturity frame w ork grounded in spatial characteristics of oil glands on pomelo peel surf aces. − A deterministic OpenCV -based pipeline inte grating preprocessing, subre gion se gmentation, morphological ltering, and rule-based decision logic. − Dual-stage v alidation combining eld-collected pomelo images with synthetic dot-pattern e xperiments to e v aluate geometric detection beha vior independently of biological v ariability . Gi v en the limited dataset and single-culti v ar scope, this w ork is positioned as a proof-of-concept baseline rather than a generalized prediction model, pro viding an engineering reference point for morphology- based surf ace analysis in lo w-cost agricultural vision systems. P omelo maturity classication fr om eld-acquir ed ima g es using oil-gland ... (Sopapun Suwansawang) Evaluation Warning : The document was created with Spire.PDF for Python.
574 ❒ ISSN: 2502-4752 2. MA TERIALS AND METHODS The proposed approach emplo ys a deterministic image-processing pipeline to quantify oil-gland mor - phology on pomelo peel surf aces and classify fruit maturity stages. The frame w ork w as designed to pro vide an interpretable, training-free solution e x ecutable on standard CPU hardw are, suitable for practical agricultural en vironments where controlled acquisition conditions and lar ge labelled datasets are not a v ailable [12]. The w orko w consists of four sequent ial stages: (i) image acquisition and preprocessing, (ii) se gmen- tation and oil-gland detection, (iii) spatial feature e xtraction and feature analysis, and (i v) rule-based maturity classication, as illustrated in Figure 1. Pomelo images were rst acquired under natural illumination and processed through grayscale con- v ersion, Gaussian smoothing, adapti v e thresholding [13], and morphological ltering [14] to enhance oil-gland structures while s upp r essing background noise. Each image w as di vided into smaller subre gions to reduce the inuence of peel curv ature during detection. Oil glands were subsequently identied using a multi-scale con v olution-based det ection strate gy [15]. Detected glands were analysed to e xtract spatial fea tures includ- ing centroid coordinates and inter -gland distances. Finally , fruit maturity w as dete rmined using a hierarchical deterministic rule deri v ed from observ ed oil-gland count statistics. All algorithms were implemented in Python 3.10 using OpenCV 4.5 for image-processing operations and NumPy for numerical computation. Experiments were e x ecuted on a laptop computer equipped with an Intel Core i7 processor . Figure 1. Proposed pipeline for pomelo maturity classication. Left: processing stages — (A) ra w image, (B) grayscale con v ersion, (C) subre gion se gmentation, (D) adapti v e thresholding and morphological ltering, (E) multi-scale oil-gland detection with centroid e xtraction, and (F) classication output with detected gland o v erlay and maturity label right: representati v e results for each stage 2.1. Image acquisition and pr epr ocessing A total of 33 Citrus maxima (Khao Y ai culti v ar) pomelo samples were collected, represent ing three maturity stages: 1–3 months, 4–5 months, and 6–8 months post-anthesis. Each maturity group comprised Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 2, May 2026: 572–583 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 575 11 samples to maintain balanced class representation. Gi v en the e xploratory nature of this feasibility study , the dataset size is consistent with pilot-scale morphological assessments reported in the literature [10]. Samples at the 1–3 and 4–5 months stages were photographed on-tree in the orchard under natural outdoor illumination. Samples at the 6–8 months stage were collected follo wing harv est — standard com- mercial practice for this culti v ar , in which mature fruit is detached to allo w post-harv est conditioning that impro v es a v our de v elopment — and photographed at an on-site storage f acility under natural ambient illumi- nation without controlled lighting or calibration panels. No controlled lighting conditions or calibration panels were emplo yed at an y stage; all images are therefore classied as eld-acquired. Images were captured using consumer -grade RGB de vices (iPhone SE and iP ad Air 9th generation, 12 MP rear cameras) [16]. Each image w as manually cropped to 600 × 600 pix els and con v erted to grayscale. Gaussian smoothing with a 5 × 5 k ernel w as applied to suppress high-frequenc y noise while preserving gland boundaries, consistent with preprocessing approaches used in citrus surf ace analysis [16]. Image binarization w as performed using adapti v e Gaussian thresholding with block size 11 and con- stant C = 1 under the THRESH BINARY INV conguration. Adapti v e thresholding w as selected o v er global thresholding [17] to accommodate local illumination v ariation across the peel surf ace [12], [13]. Morphologi- cal ltering w as subsequently applied using a 2 × 2 structuring element: erosion (tw o iterations) follo wed by dilation (one iteration), to remo v e small artef acts while preserving gland structures [10], [14]. 2.2. Segmentation and oil-gland detection T o reduce spatial v ariability introduced by peel surf ace curv ature, each 600 × 600 pix el image w as di- vided into four 300 × 300 pix el subre gions, as sho wn in Figure 1(C). Processing smaller local patches mitig ates the ef fect of une v en peel geometry on binarization consistenc y . Oil glands were detected using a multi-scale con v olution-based procedure designed to capture st ruc- tures of v arying sizes, corresponding to stage (E) in Figure 1. In the rst detection pass, con v olution k ernels of sizes 3 × 3 , 5 × 5 , and 7 × 7 were applied sequentially . The use of multiple k ernel scales follo ws the principle that surf ace structures of dif ferent sizes require corre spondingly scaled detection operators [12], [15]. Candidate re gions e xceeding empirically dened intensity thres holds were identied as gland structures using connected-component analysis [18] and mask ed to pre v ent repeated detections across k ernel scales. A second detection pass w as performed on upscaled subre gions using bilinear interpolation to impro v e sensiti vity to smaller gland structures, particularly rele v ant for distinguishing earlier maturity stages. Detected gland re gions were cate gorised into three size groups (small, medium, and lar ge) based on relati v e area ranges observ ed within the dataset. 2.3. Spatial featur e extraction P airwise Euclidean distances between detected gland centroids were computed within each subre gion, yielding a symmetric distance matrix D ∈ R N × N , where N denotes the number of detected glands. The computational comple xity of this operation is O ( N 2 ) . Mean inter -gland distance w as calculated separately for each size group. F or lar ge glands: ¯ d L = 2 n L ( n L − 1) X 1 ≤ i<j ≤ n L d i,j , (1) where n L is the number of detected lar ge glands and d i,j is the Euclidean distance between gland centroids i and j . Equi v alent formulations were applied to medium ( ¯ d M ) and small ( ¯ d S ) gland groups. The nearest-neighbour distance for each gland i w as dened as: d i, nn = min j ̸ = i d i,j . (2) together , ¯ d and d i, nn characterise global spacing and local clustering beha viour of oil glands across the peel surf ace, respecti v ely . 2.4. F eatur e analysis and thr eshold selection Both inter -gland distance descriptors ( ¯ d L , ¯ d M , ¯ d S , d i, nn ) and gland count statistics were e xamined as candidate classi cation features. Inspect ion of the collected datas et re v ealed that gland count v ariables — particularly the number of lar ge glands ( n L ) and the number of small glands in the second detection pass P omelo maturity classication fr om eld-acquir ed ima g es using oil-gland ... (Sopapun Suwansawang) Evaluation Warning : The document was created with Spire.PDF for Python.
576 ❒ ISSN: 2502-4752 ( n S ) — e xhibited more consist ent separation between maturity groups than distance-based descriptors under the present acquisition conditions. Distance features sho wed o v erlapping distrib utions across stages and were therefore not incorporated into the nal decision rule. Their computation nonetheless contrib utes to the mor - phological characterisation reported in this study and may serv e as additional discriminati v e features in future w ork with lar ger datasets. 2.5. Rule-based maturity classication The classication rule w as constructed from the tw o gland count features identied in the preceding analysis, n L and n S , on the basis of their observ ed group separation in the dataset, as illustrated in Figure 2. Primary criterion — lar ge gland count ( n L ). Figure 2(A) sho ws the distrib ution of n L across the three maturity stages. The 6–8 month group produced consistently higher counts (range: 29–178) than the 4–5 month group (range: 0–23). The threshold n ∗ L = 40 w as selected as the lo west v alue that maximises separation between the mature group and the tw o earlier stages in the collected dataset . It is noted that tw o samples in the 6–8 month group (Samples 4 and 5; n L = 29 ) fell belo w this threshold, and three samples in the 1–3 month group (Samples 6, 10, and 11; n L = 56 , 80 , and 94 ) e xceeded it. These o v erlaps are ackno wledged as a limitation of the count-based rule under the present dataset size and acquisition conditions. Secondary criterion — small gland count ( n S ). F or samples classied as non-mature under the pri- mary criterion ( n L ≤ 40 ), Figure 2(B) sho ws that n S from the second detection pass pro vided further s epara- tion between the 1–3 month group (range: 125–308) and the 4–5 month group (range: 0–317). The threshold n ∗ S = 156 w as selected from the observ ed distrib ution. One sample in the 4–5 month group (Sample 11; n S = 317 ) e xceeded this threshold, representing a potential misclassication under the se cond a ry criterion. The complete classication procedure is summarised in Algorithm 1. Figure 2. Oil gland count distrib utions across maturity stages ( n = 11 per stage, Khao Y ai culti v ar). (A) lar ge gland counts with primary threshold n ∗ L = 40 ; circled mark ers indicate samples crossing the threshold. (B) small gland counts (second detection pass) with secondary threshold n ∗ S = 156 . Thresholds are empirical and culti v ar -specic Algorithm 1 Hierarchical maturity classication rule Requir e: n L : lar ge gland count; n S : small gland count (second pass) 1: if n L > 40 then 2: Output: Mature (6–8 months) 3: else if n S < 156 then 4: Output: Intermediate (4–5 months) 5: else 6: Output: Immature (1–3 months) 7: end if The thresholds n ∗ L = 40 and n ∗ S = 156 were selected by inspection of the observ ed count distrib utions and are specic to the Khao Y ai culti v ar under the described acquisition conditions. The o v erall classication accurac y achie v ed by this rule on the collected dataset is report ed in section 3. Gi v en the limited sample size ( n = 11 per stage) and single-culti v ar scope, these thresholds represent culti v ar -specic empirical parameters of a proof-of-concept feasibility study rather than statistically optimised classication criteria. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 2, May 2026: 572–583 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 577 3. RESUL TS The pipeline w as e v aluated in tw o stages: rst on synthetic images to v alidat e detection reliability , t hen on the collected Khao Y ai pomelo dataset to assess morphological characterisation and maturity classication performance. 3.1. Algorithm v alidation on synthetic data Detection reliability w as assessed on 100 synthetically generated images prior to real-image e v aluati on (Figure 3). Each image contained precisely placed dot patterns representing 30 small, 20 medium, and 10 lar ge gland structures, distrib uted randomly to simulate natural spatial v ariability . Size-specic detection accuracies were 98.7% for lar ge, 95.9% for medium, and 95.4% for sm all glands, yielding an o v erall detection rate of 96.1%. Detection errors occurred at sites of spatial o v erlap, where adjacent structures mer ged or smaller ones were occluded by lar ger neighbours, consistent with kno wn lim- itations of morphological-lte ring approaches in dense re gions [10], [12]. These re sults conrmed suf cient detection reliability to proceed with real-image e v aluation. Figure 3. Representati v e synthetic test images used for algorithm v alidation. Each panel sho ws detected gland structures colour -coded by size: small (blue), medium (green), and lar ge (red circles). Ground-truth counts were 30 small, 20 medium, and 10 lar ge per image; detected totals ranged from 57 to 60, reecting the o v erlap-induced errors described in the te xt 3.2. Mor phological featur e characterisation Descripti v e statistics for inter -gland distance and gland count features, computed according to the spatial feature e xtract ion described in section 2, are reported for correctly classied samples in T able 2. The subset comprises n = 7 , 10 , and 9 sa mples for the 1–3, 4–5, and 6–8 month groups, respecti v ely; misclassied samples were e xcluded because their feature v alues de viated from the group-le v el distrib utions on which the rule w as deri v ed, and their inclusion w oul d conate classication-rule performance with morphological char - acterisation. The classi cation beha viour of all 33 samples, including misclassied cases, is fully documented in section 3.3 and Figure 4. P omelo maturity classication fr om eld-acquir ed ima g es using oil-gland ... (Sopapun Suwansawang) Evaluation Warning : The document was created with Spire.PDF for Python.
578 ❒ ISSN: 2502-4752 Three observ ations are notable. First, medium gland inter -gland distances were highly stable across all maturi ty stages ( C V < 5% ), indicating that medium gland spacing is structurally in v ariant during pomelo de v elopment and is ther efore not useful as a discriminati v e feature in its o wn right. Se cond , lar ge gland counts increased approximately v efold from the 1–3 month group ( 20 . 43 ± 10 . 14 ) to the 6–8 month group ( 104 . 22 ± 49 . 36 ), consistent with progressi v e oil gland dif ferentiation during citrus fruit maturation [8], [19]. Third, small gland counts from the rst detection pass declined from 2332 . 71 ± 284 . 11 to 1629 . 22 ± 204 . 51 across the same interv al, suggesting concurrent structural rem o de lling of smaller gland units, a pattern note d in broader citrus peel de v elopment studies [8]. The 4–5 month group e xhibited the highest v ariability in lar ge gland count ( C V = 124 . 06% ), reecting the biological heterogeneity e xpected during a transitional de v elopmental stage [19]. T able 2. Summary statistics (mean ± SD; CV %) of inter -gland dis tances and oil gland counts for correctly classied samples across maturity stages ( n = 7 , 10 , and 9 for the 1–3, 4–5, and 6–8 month groups, respecti v ely). Distance v alues are in pix els Feature 1–3 months 4–5 months 6–8 months Lar ge dist. (px) 219 . 63 ± 67 . 31 (30.65) 98 . 50 ± 128 . 20 (130.15) 128 . 40 ± 41 . 46 (32.29) Medium dist. (px) 54 . 12 ± 2 . 44 (4.51) 54 . 59 ± 1 . 84 (3.38) 54 . 22 ± 2 . 01 (3.71) Small dist. (px) 30 . 65 ± 2 . 34 (7.63) 33 . 06 ± 1 . 0 6 (3.20) 3 2 . 76 ± 1 . 81 (5.53) Lar ge count 20 . 43 ± 10 . 14 (49.59) 5 . 60 ± 6 . 94 (124.06) 1 04 . 22 ± 49 . 36 (47.35) Medium count 508 . 43 ± 45 . 16 (8.88) 474 . 7 0 ± 138 . 51 (29.18) 640 . 67 ± 56 . 57 (8.83) Small count (1st pass) 2332 . 71 ± 284 . 11 (12.18) 2262 . 80 ± 2 77 . 19 (12.25) 1629 . 22 ± 204 . 51 (12.55) Small count (2nd pass) 191 . 71 ± 21 . 58 (11.26) 109 . 20 ± 36 . 71 (33.62) 227 . 33 ± 83 . 06 (36.54) 3.3. Classication perf ormance Applying the hierarchical rule (Algorithm 1) to all 33 collected samples yielded an o v erall accurac y of 78.79% (26/33). Class-wise results, deri v ed from the confusion matrix in Figure 4, are summarised in T able 3. Figure 4. Left: confusion matrix for the hierarchical rule-based classier applied to 33 Khao Y ai pomelo samples (o v erall accurac y: 78.79%). Right: precision, recall, and one-vs-all accurac y per maturity group T able 3. Class-wise e v aluation metrics deri v ed from the confusion matrix in Figure 4. The macro-a v eraged F1-score is reported in the nal ro w as the unweighted mean of the three class-wise F1-scores [20], [21] Group TP FP FN Prec. (%) Rec. (%) F1 1–3 months 7 2 4 77.78 63.64 0.700 4–5 months 10 2 1 83.33 90.91 0.870 6–8 months 9 3 2 75.00 81.82 0.783 Macro a v erage 0.784 Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 2, May 2026: 572–583 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 579 The macro-a v eraged F1-score of 0.784 w as computed as the unweighted mean of the three class-wi se F1-scores follo wing the formulation in Sok olo v a and Lapalme [20]. Since the present dataset i s class-balanced ( n = 11 per stage), the macro-a v eraged and weighted F1-scores are numerically equi v ale n t ; the macro a v erage is reported as it treats each maturity class equally re g ardless of class size [20], [21]. F or balanced multiclass datasets, the F1-score has been sho wn to pro vide a more informati v e summary of classier performance than o v erall accurac y alone [20], [22]. Precision, recall, and F1-score were computed from the per -class confusion matrix entries using standard denitions [20]. The 4–5 month group achie v ed the highest F1-score (0.870) and recall (90.91%), reecting strong separation under the secondary criterion n ∗ S = 156 . The 1–3 month group yielded the lo west recall (63.64%), attrib utable to the three samples (S6, S10, and S11) whose lar ge gland counts e xceeded the primary threshold n ∗ L = 40 , as documented in section 2. Importantly , no sample from the 1–3 month group w as misclassied as mature (6–8 months), nor vice v ersa: all misclassications in v olv ed adjacent maturity stages only . This result indicates that the pipeline correctly captures the lar gest de v elopmental contrast between immature and mature fruit, consistent with the progressi v e oil gland dif ferentiation reported in citrus peel studies [8], [10]. 3.4. Computational cost The mean processing time per image w as 57 seconds, m easured on an Inte l Core i 7 l aptop computer (Python 3.10, OpenCV 4.5, single-threaded CPU e x ecution, n = 33 samples), encompassing the complete pipeline from preprocessing to classication output. No GPU acceleration w as emplo yed. The processing time is primarily attrib utable to the iterati v e st ructure of the detection stage, which applies con v olution k ernels of three scales ( 3 × 3 , 5 × 5 , and 7 × 7 ) s equentially in the rst pass, follo wed by a second detection pass on upscaled subre gions to capture smaller gland structures. Each pass in v olv es per - subre gion masking and candidate v erication, resulting i n repeated tra v ersals o v er the image content before a classication decision is reached. This deliberate multi-pass design prioritises detection completeness o v er processing speed, consistent with the proof-of-concept objecti v es of this study . While the curr ent implementation precludes real-time eld deplo yment, the pipeline requires no model training, labelled datasets, or specialist inference infrastructure. Computational ef cienc y could be impro v ed in future w ork through k ernel parallelisation or replacing the e xhausti v e multi-pass strate gy with a single-pass learned detector , without altering the interpretable rule-based classication stage [12]. 3.5. Statistical v alidation The consistenc y of group-le v el performance w as assessed using the one-sided e xact binomial test [23], with the null h ypothesis that each group’ s true cla ssication accurac y is no lo wer than the o v erall observ ed accurac y of the full dataset ( p 0 = 0 . 7879 ): H 0 : p ≥ p 0 , H 1 : p < p 0 . (3) This formulation tests whether an y indi vidual group performs signicantly w orse than the data set- le v el a v erage. p -v alues were computed from the e xact binomial probability mas s function; tw o-sided Clopper – Pearson 95% condence interv als were reported as a complementary measure of uncertainty [23]. Results are presented in T able 4. No group returned a p -v alue belo w α = 0 . 05 , indicating that the null h ypothesis w as not rejected for an y group. The 1–3 month group sho wed the widest condence interv al [0.308, 0.891], reecting the lo wer number of correct classications ( k = 7 of 11) combined with the small per - group sample size. These results are consistent with stable group-le v el performance within the present dataset; the y should not be interpreted as e vidence of generalisation be yo nd the Khao Y ai culti v ar or the described acquisition conditions, gi v en the limited sample size and single-culti v ar scope [10]. T able 4. One-sided e xact binomial test results per maturity group. H 0 : group accurac y ≥ p 0 = 0 . 7879 ; α = 0 . 05 . T w o-sided Clopper –Pearson 95 % CIs are sho wn Group Lo wer CI Upper CI p -v alue 1–3 months 0.308 0.891 0.189 4–5 months 0.587 0.998 0.927 6–8 months 0.482 0.977 0.712 P omelo maturity classication fr om eld-acquir ed ima g es using oil-gland ... (Sopapun Suwansawang) Evaluation Warning : The document was created with Spire.PDF for Python.
580 ❒ ISSN: 2502-4752 4. DISCUSSION 4.1. Inter pr etation of classication perf ormance The o v erall accurac y of 78.79% and macro-a v eraged F1-score of 0.784 are consistent with the e x- ploratory scope of this study . Al l misclassications occurred between adjacent maturity stages; no sample from the 1–3 month group w as misclassied as mature, nor vice v ersa. This result indicates that the pipeline correctly captures the most pronounced de v elopmental contrast in oil-gland morphology , consistent wi th the progressi v e gland dif ferentiation described in citrus peel studies [8], [10]. The lo wer recall of the 1–3 month group (63.64%) reects the o v erlap in lar ge gland counts between early-stage and transitional samples, which is a kno wn source of dif culty in rule-based maturity assessment under biological v ariability [19]. The intermediate group (4–5 months) achie v ed the highest F1-score (0.870), cons istent with the dis- tinct morphological chara cteristics observ ed during the transiti onal de v elopmental phase. The wide condence interv al for the 1–3 month group [0.308, 0.891] reects both the lo wer classication accurac y and the inherent uncertainty associated with small per -group sample sizes [23]. These results should be i nterpreted within the constraints of a proof-of-concept study rather than as e vidence of generalisation be yond the present culti v ar and acquisition conditions. 4.2. Contrib utions r elati v e to prior w ork This study presents a transparent, rul e-based pipeline that quanties oil-gland count and spatial mor - phology from eld-acquired images without requiring model training, labelled datasets, or specialist sensing equipment. In cont rast to spectroscopic and acoustic methods [1], the propos ed approach relies solely on consumer -grade RGB imaging under natural illumination. Compared with deep learning classiers [4], [6], the rule-based frame w ork of fers e xplicit decision traceability at the cost of reduced predicti v e capacity . Medium gland inter -gland distances demonstrated e xceptional stability across all maturity stages ( C V < 5% ), suggesting that medium gland spacing is structurally in v ariant during pomelo de v elopment. This nding is consistent with the de v elopmental biology of citrus a v edo tissue reported in the literature [8] and represents a no v el morphometric observ ation for the Khao Y ai culti v ar . The synthetic dot-pattern v ali- dation, which achie v ed an o v erall detection rate of 96.1%, pro vides a geometry-independent baseline for the detection algorithm that is decoupled from biological v ariability . This dual-stage v alidation design strengthens condence in algorithmic beha viour prior to real-sample deplo yment [10], [12]. 4.3. Limitations Se v eral limitations constrain the scope of the present ndings: − Dataset size and culti v ar scope. The dataset of 33 samples from a single culti v ar (Khao Y ai) limits statistical po wer and generalisability . The deri v ed thresholds ( n ∗ L = 40 , n ∗ S = 156 ) are culti v ar -specic empirical parameters rather than uni v ersal classication criteria. − Illumination v ariability . Acquisition heterogeneity . Images of the 1–3 and 4–5 month groups were captured on-tree in the orchard, whereas 6–8 month samples were photographed post-harv est at an on-site storage f acility . Although no controlled lighting w as emplo yed in either setting, the dif ference in acquisition con- te xt may introduce systematic v ariation in background te xture and illumination angle that cannot be fully disentangled from maturity-related morphological dif ferences in the present dataset. − Manual preprocessing. Images required manual cropping to 600 × 600 pix els. This step introduces operator dependenc y and limits throughput in practical deplo yments. − Processing time. The mean processing time of 57 seconds per image precludes real-time eld deplo yment in the current implementation, as discussed in section 3. − 2D projection. Analysis of the curv ed peel surf ace from a 2D image introduces potent ial geometric distor - tion in density and distance measurements, particularly for samples with pronounced surf ace curv ature. These constraints reect common challenges in transitioning image-based sensing from controlled laboratory conditions to practical orchard en vironments [1], [16]. 4.4. Futur e w ork Three directions are prioritised for future in v estig ation. Fi rst, dataset e xpansion across multiple Khao Y ai orchards and additional Citrus maxima culti v ars is required to assess the generalisability of the de- ri v ed thresholds and to support statistically rob ust threshold optimisation. Point-pattern diagnostic approaches could be applied to characterise inter -gland spatial distrib utions more rigorously across culti v ars [24]. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 2, May 2026: 572–583 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 581 Second, automated re gion-of-interest e xtraction through semantic se gmentation w ould eliminate the current manual cropping dependenc y , enabling higher throughput and reducing operator v ariability [25]. In- te gration of illumination normalisati on or colour constanc y preprocessing could further impro v e binarization rob ustness under v ariable eld conditions [16]. Third, computational optimisation of the multi-scale detect ion stage — the primary bottleneck at the current processing rate of 57 seconds per image — represents a necessary step to w ard practical deplo yment. K ernel parallelisation or replacement of the e xhausti v e multi-pass strate gy with a single-pass learned detector could substantially reduce processing time. Deplo yment on dedicated embedded vision hardw are has been demonstrated as feas ible for comparable citrus sorting applications [26], pro viding a practical deplo yment pathw ay for a rule-based pipeline of the type proposed here. 5. CONCLUSION This study in v estig ated the feasibility of oil-gland morphology as an interpretable, non-dest ructi v e maturity indicator for pomelo ( Citrus maxima , Khao Y ai culti v ar) using a rule-based image-processing pipeline. The pipeline achie v ed an o v erall ac curac y of 78.79% and a macro-a v eraged F 1-score of 0.784 on a balanced dataset of 33 eld-acquired samples ( n = 11 per stage). Synthetic dot-pattern v alidation conrmed a detection rate of 96.1%, establishing algorithmic reliability independently of biological v ariability . No misclassication occurred between the most de v elopmentally distinct groups (1–3 and 6–8 months), indicating that the pipeline correctly captures the principal oil-gland de v elopmental contrast during pomelo maturation. The frame w ork requires no model training, labelled datasets, or specialist imagi ng hardw are, operating on consumer -grade RGB images captured under natural illumination. These chara cteristics distinguish it from spectroscopic and deep-learning approaches that depend on controlled conditions or e xtensi v e training data. The ndings are interpreted as methodological feasibility rather than v alidated generalisation. The empirical thresholds and morphological patterns reported here are specic to the Khao Y ai culti v ar under the described acquisition conditions and establish a quantitati v e basel ine for future multi-culti v ar , multi-orchard in v estig ations of morphology-based maturity assessment in pomelo. A CKNO WLEDGMENTS The authors w ould lik e to e xpress their sincere gratitude to Kiattisak Naria, for his dedicated contri- b ution as a research assistant. FUNDING INFORMA TION This w ork w as supported by the Adv anced Signal Processing for Disrupti v e Inno v ation Research Center , Nakhon P athom Rajabhat Uni v ersity . A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contri b ut or Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Sopapun Suw ansa w ang ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Harutai Dinsakul ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ W irot Buang am ✓ ✓ ✓ ✓ Jiraroj T osasukul ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V isualization M : M ethodology R : R esources Su : S upervision So : S oftw are D : D ata Curation P : P roject Administration V a : V alidation O : Writing - O riginal Draft Fu : F unding Acquisition F o : F ormal Analysis E : Writing - Re vie w & E diting P omelo maturity classication fr om eld-acquir ed ima g es using oil-gland ... (Sopapun Suwansawang) Evaluation Warning : The document was created with Spire.PDF for Python.